Classifications

• • G—PHYSICS
  • G06—COMPUTING; CALCULATING OR COUNTING
  • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
  • G06T7/00—Image analysis
  • G06T7/20—Analysis of motion
• • G—PHYSICS
  • G06—COMPUTING; CALCULATING OR COUNTING
  • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
  • G06T7/00—Image analysis
  • G06T7/30—Determination of transform parameters for the alignment of images, i.e. image registration
  • G06T7/38—Registration of image sequences
• • G—PHYSICS
  • G06—COMPUTING; CALCULATING OR COUNTING
  • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
  • G06T2207/00—Indexing scheme for image analysis or image enhancement
  • G06T2207/10—Image acquisition modality
  • G06T2207/10016—Video; Image sequence
• • G—PHYSICS
  • G06—COMPUTING; CALCULATING OR COUNTING
  • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
  • G06T2207/00—Indexing scheme for image analysis or image enhancement
  • G06T2207/30—Subject of image; Context of image processing
  • G06T2207/30232—Surveillance

Definitions

• the present invention relates to methodologies and systems for combining and integrating visual information from various sources.
• a sequence of images contains much more information than any individual image does.
• a sequence of images of a scene contains temporal changes caused by dynamic changes in the scene, or temporal changes induced by the motion of the camera. This information is employed by the present invention to provide alignment between images of the two sequences of images even when the sequences have no spatial overlap or have very low similarity.
• This invention seeks to provide methodologies and systems for combining and integrating visual information from various sources.
• the present invention replaces a requirement of "coherent appearance", which is a fundamental assumption in prior art image alignment methods, with a requirement of "consistent temporal behavior".
• the requirement of "consistent temporal behavior” is easier to satisfy, for example by producing two sequences of images using two cameras moved jointly in space.
• the present invention is therefore useful not only in cases of non-overlapping sequences, but also in cases which are inherently difficult for standard image alignment techniques.
• the two cameras are attached closely to each- other, so that their centers of projection are very close, and move jointly in space.
• the induced frame-to-frame transformations within each sequence have -correlated behavior across the two sequences.
• a preferred embodiment of the present invention employs correlated temporal behavior to resolve both spatial and temporal transformations between images of the two sequences.
• the present invention is therefore, useful for a- variety of real-world applications, including:
• Multi-sensor alignment for image fusion This requires accurate alignment of images obtained by sensors of different sensing modalities, such as Infra-Red and visible light. Such images differ significantly in their appearance due to different sensor properties.
• correlated includes inter alia the following: mathematical correlation, statistical correlation and any other co-relationship.
• a method for computing at least one of spatial and temporal relationships between at least first and second sequences of representations having respective first and second temporal progressions includes employing the first and second temporal progressions to obtain at least one of the spatial and temporal relationships.
• a method for computing at least one of spatial and temporal relationships between at least first and second sequences of representations having respective first and second temporal progressions includes employing only the first and second temporal progressions to obtain at least one of the spatial and temporal relationships.
• a system for computing at least one of spatial and temporal relationships between at least first and second sequences of representations having respective first and second temporal progressions includes input functionality for receiving the first sequence of representations and the first temporal progression and the second sequence of representations and the second temporal progression and computation functionality employing at least one of the first sequence of representations and the first temporal progression and at least one of the second sequence of representations and the second temporal progression to obtain at least one of the spatial and temporal relationships.
• a system for computing at least one of spatial and temporal relationships between at least first and second sequences of representations having respective first and second temporal progressions includes input functionality for receiving the first and second temporal progressions and computational functionality employing the first and second temporal progressions to obtain at least one of the spatial and temporal relationships.
• the method also includes computing the first and second temporal progressions which are then employed to obtain at least one of the spatial and temporal relationships.
• the representations include visual representations, images and at least three-dimensional objects. Still further in accordance with a preferred embodiment of the present invention the step of employing obtains the spatial relationship, and/or the temporal relationship.
• the employing step obtains the spatial and temporal relationships in the absence of common spatial information between the representations. Further in accordance with a preferred embodiment of the present invention the step of employing obtains the spatial and temporal relationships in the absence of common information between individual ones of the representations belonging to different ones of the at least first and second sequences.
• the step of employing obtains the spatial and temporal relationships in the absence of common information between any individual ones of the representations belongingto different ones of the at least first and second sequences.
• the spatial relationship includes at least one parameter of a geometric transformation between the at least first and second sequences of representations.
• the spatial relationship includes a 2-dimensional projective transformation.
• the spatial relationship includes a 3-dimensional projective transformation. Further in accordance with a preferred embodiment of the present invention the spatial relationship includes a fundamental matrix. -
• the spatial relationship includes a 2-dimensional parametric transformation.
• the spatial relationship includes an at least 3-dimensional parametric transformation.
• the spatial relationship includes a 2-dimensional non-parametric transformation.
• the spatial relationship includes an at least 3-dimensional non-parametric transformation.
• the spatial relationship includes a spatial relationship between individual ones of the representations belonging to different ones of the at least first and second sequences.
• the spatial relationship includes a spatial relationship between individual ones of the representations belonging to different ones of the at ' least first and second sequences, at least one of the individual ones of the representations being an interpolated representation.
• the spatial relationship includes a spatial relationship between individual ones of the representations belonging to different ones of the at least first and second sequences, one of the individual ones of the representations being an interpolated representation and the other of the individual ones of the representations being a non-interpolated representation.
• the temporal relationship includes at least one parameter of a temporal transformation between the at least first and second sequences of representations.
• the temporal relationship includes a time shift between the at least first and second sequences of representations.
• the temporal relationship includes an affine transformation in time, a parametric transformation in time and/or a non-parametric transformation in time.
• the temporal relationship includes a temporal relationship in time between individual ones of the representations belonging to different ones of the at least first and second sequences.
• the temporal relationship includes a temporal relationship in time between individual ones of the representations belonging to different ones of the at least first and second sequences.
• one of the individual ones of the representations is an interpolated representation and the other of the individual ones of the representations is a non-interpolated representation.
• the interpolated representation is interpolated in time and the interpolated representation is interpolated in space.
• first and second temporal progressions include ordered intra-sequence representation-to-representation transformations. Still further in accordance with a preferred embodiment of the present invention the first and second temporal progressions include ordered intra-sequence representation-to-representation transformations resulting from relative motion between sensors and a scene.
• the intra-sequence representation-to-representation transformations include 2-dimensional projective transformations, and/or 3-dimensional projective transformations.
• the intra-sequence representation-to-representation transformations include fundamental matrices, 2_-dimensional parametric transformations, at least 3-dimensional parametric transformations,. 2-dimensional non-parametric transformations, at least 3-dimensional non-parametric transformations and camera matrices.
• the step of employing includes correlating the first and second temporal progressions. Additionally in accordance with a preferred embodiment of the present invention the step of employing includes equating properties of the first temporal progression and the second temporal progression. Further in accordance with a preferred embodiment of the present invention the step of employing includes correlating properties of the first temporal progression and the second temporal progression.
• the step of employing includes equating a sequential application and at least one of the intra-sequence representation-to-representation transformations of the first temporal progression and an unknown the spatial relationship between the at least first and second sequences with a sequential application of the unknown spatial relationship between the at least first and second sequences and at least one of the intra-sequence representation-to-representation transformations of the second temporal progression.
• the step of employing includes equating a composition and at least one of the intra-sequence representation-to-representation transformations of the first temporal progression and an unknown the spatial relationship between the at least first and second sequences with a composition of the unknown spatial relationship between the at least first and second sequences and at least one of the intra-sequence representation-to-representation transformations of the second temporal progression.
• the step of employing includes obtaining an unknown the spatial relationship between the at least first and second sequences by equating a sequential application and at least one of the intra-sequence representation-to-representation transformations of the first temporal progression and the unknown spatial relationship between the at least first and second sequences with a sequential application of the unknown spatial relationship between the at least first and second sequences and at least one of the intra-sequence representation-to-representation transformations of the second temporal progression.
• the step of equating includes equating up to a scale factor.
• intra-sequence representation-to-representation transformations include multiple simple motions taking place at least partially at different times.
• the intra-sequence representation-to-representation transformations include multiple combinations of multiple simple motions taking place at least partially at different times.
• the intra-sequence representation-to-representation transformations include multiple complex motions taking place at least partially at different times.
• the intra-sequence representation-to-representation transformations include multiple combinations of multiple complex motions taking place at least partially at different times.
• first and second temporal progressions include ordered intra-sequence representation-to-representation transformations at least some of which result from relative motion between sensors and a scene.
• the step of employing uses multiple combinations of intra-sequence representation-to-representation transformations.
• the spatial relationship between the at least first and second sequences of representations results from an acquisition relationship, between first and second sensors acquiring respective the at least first and second sequences, being fixed over time.
• the spatial relationship between the at least first and second sequences of representations results from an acquisition relationship, between first and second sensors acquiring respective the at least first and second sequences, changes in a known way over time.
• the acquisition relationship includes relative position, relative orientation and relative internal sensor parameters.
• the acquisition relationship is not known.
• the at least first and second sequences are acquired generally at the same time.
• the at least first and second sequences are acquired generally at different times. Further in accordance with a preferred embodiment of the present invention the at least first and second sequences represent measurements from the same scene.
• the at least first and second sequences represent measurements from different portions of the same scene.
• the at least first and second sequences represent measurements from different overlapping -portions of the same scene.
• the at least first and second sequences represent measurements from different non-overlapping portions of the same scene.
• the at least first and second sequences represent measurements from different scenes.
• the scene is two-dimensional.
• the scene is at least three-dimensional.
• the scene is static.
• the scene is dynamic.
• the measurements are generally the same for each sensor. Still further in accordance with a preferred embodiment of the present, invention the measurements are generally different for each sensor.
• the measurements include at least one of illumination, heat, radiance, electromagnetic radiation, color, distance, density, sound and speed.
• the method also includes the step of employing at least one of the spatial and temporal relationships for sequence fusion.
• the method also includes step of employing at least one of the spatial and temporal relationships for alignment of sequences obtained at different zooms. Still further in accordance with a preferred embodiment of the present invention the method also includes the step of employing at least one of the spatial and temporal relationships for surveillance. Further in accordance with a preferred embodiment of the present invention the method also includes the step of employing at least one of the spatial and temporal relationships for generation of wide-screen movies from multiple at least partially non-overlapping narrow field of view movies. Additionally in accordance with a preferred embodiment of the present invention the method also includes the step of employing at least one of the spatial and temporal relationships for image fusion.
• the method also includes the step of employing at least one of the spatial and temporal relationships for integrating information contained in the at least first and second sequences.
• the method also includes the step of employing at least one of the spatial and temporal relationships for alignment of images obtained at different zooms. Additionally in accordance with- a preferred embodiment of the present ' invention the method also includes the step of employing at least one of the spatial and temporal relationships for comparing information contained in the at least first and second sequences.
• the method also includes the step of employing at least one of the spatial and temporal relationships for finding differences between information contained in the at least first and second sequences.
• the method also includes the step of employing at least one of the spatial and temporal relationships for finding differences between information contained in the at least first and second sequences relating to the same scene at different times.
• the method " also includes the step of employing at least one of the spatial and temporal relationships for integrating information contained in the at least first and second sequences and thereby providing an information output which exceeds limitations of individual sensors.
• the step of employing includes comparing properties of the first temporal progression and the second temporal progression.
• the computation functionality includes functionality for computing the first and second temporal progressions which are then employed to obtain at least one of the spatial and temporal relationships.
• system also includes temporal progression computation functionality for computing the first and second temporal progressions- and supplying them to the input functionality.
• Fig. 1 A is a simplified illustration of two cameras, fixed to each other, each taking a sequence of images of a portion of a scene, the portions of the scene photographed by the two cameras being non-overlapping;
• Fig. IB illustrates inherent ambiguities in the relative spatial relationship of images in the sequences taken as shown in Fig. 1 A;
• Fig. IC illustrates portions of two sequences of images taken as shown. in
• Fig. ID illustrates pairs of images forming part of the sequences of images taken as shown in Fig. IA, whose spatial and temporal relationships are determined in accordance with the present invention
• Fig. 2A is a simplified illustration of two cameras, fixed to each other, each taking a sequence of images of a portion of a scene, the portions of the scene being photographed at significantly different zooms by the two cameras;
• Fig. 2B illustrates inherent ambiguities in the relative spatial relationship of images in the sequences taken as shown in Fig. 2A;
• Fig. 2C illustrates portions of two sequences of images taken as shown in
• Fig. 2D illustrates pairs of images forming part of the sequences of images taken as shown in Fig. 2A, whose spatial and temporal relationships are determined in accordance with the present invention
• Fig. 3 A is a simplified illustration of two sensors, fixed to each other, each taking a sequence of images of a portion of a scene, the portions of the scene being imaged by the two sensors employing different sensing modalities;
• Fig. 3B illustrates inherent ambiguities in the relative spatial relationship of images in the sequences taken as shown in Fig. 3A;
• Fig. 3C illustrates portions of two sequences of images taken as shown in Fig. 3 A and the unknown temporal relationship between the sequences
• Fig. 3D illustrates pairs of images forming part of the sequences of images taken as shown in Fig. 3A, whose spatial and temporal relationships are determined in accordance with the present invention
• Fig. 4A is a simplified illustration of a scene being photographed at two different times, producing two corresponding sequences of images, the path traveled by a camera used for photographing being generally identical for both sequences;
• Fig. 4B illustrates inherent ambiguities in the relative spatial relationship of images in the sequences taken as shown in Fig. 4A;
• Fig. 4C illustrates portions of two sequences of images taken as shown in Fig. 4A and the unknown temporal relationship between the sequences
• Fig. 4D illustrates pairs of images forming part of the sequences of images taken as shown in Fig. 4A, whose spatial and temporal relationships are determined in accordance with the present invention
• Fig. 5A illustrates the relationships between image to image transformations within two sequences, induced by motion of two cameras along an axis, the two cameras being arranged at 180 degrees with respect to each other along the axis;
• Fig. 5B illustrates the relationships between image to image transformations within two sequences, induced by motion of two cameras along an axis, the two cameras being arranged at 90 degrees with respect to each other, one of the cameras being aligned along the axis of motion
• Fig. 5C illustrates the relationships between image to image transformations within two sequences, induced by motion of two cameras along an axis, the two cameras being directed in the same direction perpendicular to the direction of motion but at different zooms;
• Figs. 6A, 6B and 6C together illustrate that employing one type of transformation may not be sufficient for resolving ambiguities in the spatial and temporal relationships between sequences but that employing multiple different types of transformations reduces ambiguities in the spatial and temporal relationships between sequences;
• Fig. 7 is a simplified illustration of a complex motion of two cameras, fixed to each other, each taking a sequence of images of a portion of a scene, the portions of the scene photographed by the two cameras being non-overlapping;
• Fig. 8 is a simplified illustration portions of two sequences of images taken as shown in Fig. 7, wherein the two sequences are spatially related by a fixed and unknown inter-camera homography and temporally related by a fixed and unknown time shift.
• Fig. 9 is a simplified functional block diagram of a preferred process of creating two sequences of images and aligning them.
• Figs. IA - ID illustrate resolution of spatial and temporal relationships between sequences of images taken by two moving cameras fixed to each other in accordance with the present invention.
• two cameras designated respectively by reference numerals 100 and 102 are fixed to each other in any suitable manner.
• Each camera takes a sequence of images of a portion of a scene as the cameras move while they are fixed together.
• the movement of the cameras may be any suitable movement, such as rotation and/or translation in one or more dimensions and relative to any suitable point.
• the two cameras 100 and 102 may rotate about the optical axis of one of the cameras or about any other axis.
• translation of the cameras may occur in any suitable direction.
• FIGs. IA - ID portions of the scene photographed by the two cameras are non-overlapping and are designated respectively by reference numerals 104 and 106.
• Fig. IB it is seen that the portions 104 and 106 of the scene of Fig. IA may be represented by corresponding images 108 and 110.
• images 108 and 110 each belong to a sequence of images, each produced by one of the moving cameras 100 and 102, the respective sequences being designated by reference numerals 1 12 and 114.
• sequence 112 also includes images 122, 124 and 126.
• sequence 114 also includes images 130, 132 and 134.
• a problem addressed by the present invention is that the visual- information contained in individual pairs of images, one belonging to sequence 112 and the other belonging to sequence 114, (e.g. (108, 110), (122, 130) or (122, 134)), is sufficient to establish neither the spatial nor the temporal relationship between two images of a pair. More generally, the visual information contained in individual pairs of images, one of the pair, belonging to sequence 112 and the other of the pair belonging to sequence 114, is sufficient to establish neither the spatial nor the temporal relationship between the two sequences 112 and 114.
• the unknown spatial relationship of images 108 and 110 is seen graphically by considering three examples of possible relative spatial relationships shown in Fig. IB and designated by reference numerals 116, 118 and 120. In each example, the two images 108 and 1 10 are placed, in a different spatial relationship, all of which are consistent with the visual content of the images 108 and 110.
• the unknown temporal relationship of sequences 112 and 114 is seen graphically by considering Fig. IC. It is appreciated that it is not apparent from a cursory examination of sequences 112 and 114, which images in sequence 112 are taken at the same time as which images in sequence 114.
• the present invention provides a system and technique for determining the correct relationship between the images 108 and 110 and more generally the correct spatial and temporal relationships between sequences 112 and 114, as shown in Fig. ID.
• the present invention determines which image in sequence 1 12 corresponds in time with which image in sequence 114 and further determines the spatial relationship between the images which correspond in time.
• image B (122) in sequence 112 is found to correspond in time with image c
• the present invention employs an appreciation that despite there being no common static visual information in the two sequences, the image to image dynamics in each of the two sequences are nevertheless related both spatially and temporally.
• the present invention utilizes the relationship between the dynamics to correlate the- two sequences in time and space.
• the image to image dynamics preferably are expressed by spatial transformations over time between images in each sequence.
• the problem of aligning sequences of images in time and space is thus reduced to a problem of determining the relationship between sequences of such transformations in time and in space.
• FIGs. 2A - 2D illustrate the resolution of spatial and temporal relationships between sequences of images taken by two moving cameras fixed to each other in accordance with the present invention, the portions of the scene being photographed at significantly different zooms by the two cameras.
• two cameras are fixed to each other in any suitable manner.
• Each camera takes a sequence of images of a portion of a scene as the cameras move while they are fixed together.
• the movement of the cameras may be any suitable movement, such as rotation and/or translation in one or more dimensions and relative to any suitable point.
• the two cameras may rotate about the optical axis of one of the cameras or about any other axis.
• translation of the cameras may occur in any suitable direction.
• FIGs. 2 A - 2D portions of the scene photographed by the two cameras are imaged at significantly different zooms, thus imaging various features in the scene at different spatial scales.
• Fig. 2B it is seen that the portions 204 and 206 of the scene of Fig. 2 A may be represented by corresponding images 208 and 210, respectively.
• images 208 and 210 each belong to a sequence of images, each produced by one of the moving cameras 200 and 202, the respective sequences being designated by reference numerals 212 and 214.
• sequence 212 also includes images 222, 224 and 226.
• sequence 214 also includes images 230,
• features imaged in the zoomed-in image 210 may be different from the features imaged in the zoomed-out image 208 due to the different spatial scales employed by the cameras.
• features imaged in the zoomed-in image 210 may be different from the features imaged in the zoomed-out image 208 due to the different spatial scales employed by the cameras.
• individual leaves of a tree may be discerned in image 210, while such leaves may not be discernible in image 208.
• a problem addressed by the present invention is that the visual information contained in individual pairs of images, one belonging to sequence 212 and the other belonging to sequence 214, (e.g. (208, 210), (222, 230) or (222, 234)), is sufficient to establish neither the spatial nor the temporal relationship between two images of a pair.
• the visual information contained in individual pairs of images, one of the pair belonging to sequence 212 and the other of the pair belonging to sequence 214, is sufficient to establish neither the spatial nor the temporal relationship between the two sequences 212 and 214.
• the unknown spatial relationship of images 208 and 210 is seen graphically by considering three examples of possible relative spatial relationships shown in Fig. 2B and designated by reference numerals 216, 218 and 220. In each example, the two images 208 and 210 are placed in a different spatial relationship with each other, all of which are consistent with the visual content of the images 208 and 210.
• sequences 212 and 214 are seen graphically by considering Fig. 2C. It is appreciated that it is not apparent from a cursory examination of sequences 212 and 214, which images in sequence 212 are taken at the same time as which images in sequence 214.
• the present invention provides a system and technique for determining the correct relationship between the images 208 and 210 and more generally the correct spatial and temporal relationships between sequences 212 and 214, as shown in Fig. 2D.
• the present invention determines which image in sequence 212 corresponds in time with which image in sequence 214 and further determines the spatial relationship between the images which correspond in time.
• FIG. 2D image A (208) in sequence 212 is found to correspond in time with image b (210) in sequence 214.
• the correct spatial relationship between images A (208) and b (210) is shown in Fig. 2D at reference numeral 250.
• image B (222) in sequence 212 is found to correspond in time with image c (232) in sequence 214
• image C (224) in sequence 212 is found to correspond in time with image d (234) in sequence 214.
• the correct spatial relationship between images B (222) and c (232) is shown in Fig. 2D at reference numeral 252 and the correct spatial relationship between images C (224) and d (234) is shown in Fig. 2D at reference numeral 254.
• the present invention employs an appreciation that despite there not necessarily being any common static visual information in the two sequences, the image to image dynamics in each of the two sequences are nevertheless related both spatially and temporally.
• the present invention utilizes the relationship between the dynamics to correlate the two sequences in time and space.
• the image to image dynamics preferably are expressed by spatial transformations over time between images in each sequence.
• the problem of aligning sequences of images in time and space is thus reduced to a problem of determining the relationship between sequences of such transformations in time and in space.
• FIG. 3 A - 3D illustrate resolution of spatial and temporal relationships between sequences of images taken by two moving sensors fixed to each other in accordance with the present invention, the portions of the scene being imaged by the two sensors employing different sensing modalities.
• two sensors such as an infra red (IR) camera 300 and a visible light (VL) camera 302 are fixed to each other in any suitable manner.
• Each sensor takes a sequence of images of a portion of a scene as the sensors move while they are fixed together.
• the movement of the sensors may be any suitable movement, such as rotation and/or translation in one or more dimensions and relative to any suitable point.
• the two sensors may rotate about the optical axis of one of the sensors or about any other axis.
• translation of the sensors may occur in any suitable direction.
• FIGs. 3 A - 3D portions of the scene sensed by the two sensors are imaged at different wavelength ranges, thus imaging various features in the scene in different modalities.
• the portions 304 and 306 of the scene of Fig. 3 A may be represented by corresponding images 308 and 310.
• images 308 and 310 each belong to a sequence of images, each produced by one of the moving sensors 300 and 302, the respective sequences being designated by reference numerals 312 and 314.
• sequence 312 also includes images 322, 324 and 326.
• sequence 314 also includes images 330, 332 and 334.
• features imaged in the infra-red (IR) image 308 may be different from the features imaged in the visible light (VL) image 3.10 due to the different sensing modalities employed by the sensors.
• hot regions may be discerned in image 308, while heat is not sensed in image 310.
• the visual appearance of a bus may be discerned in image 310, while features appearing only in visible light are not captured in image 308.
• a problem addressed by the present invention is that the visual information contained in individual pairs of images, one belonging to sequence 312 and the other belonging to sequence 314, (e.g. (308, 310), (322, 330) or (322, 334)), is sufficient to establish neither the spatial nor the temporal relationship between two images of a pair. More generally, the visual information contained in individual pairs of images, one of the pair belonging to sequence 312 and the other of the pair belonging to sequence 314, is sufficient to establish neither the spatial nor the temporal relationship between the two sequences 312 and 314.
• Figs. 3A - 3D two images are the to be in the same "temporal relationship" if they are taken at the same time.
• the unknown spatial relationship of images 308 and 310 is seen graphically by considering three examples of possible relative spatial relationships shown in Fig. 3B and designated by reference numerals 316, 318 and 320. In each example, the two images 308 and 310 are placed in a different spatial relationship, all of which are consistent with the visual content of the images 308 and 310.
• the unknown temporal relationship of sequences 312 and 314 is seen graphically by considering Fig. 3C. It is appreciated . that it is not apparent from a cursory examination of sequences 312 and 314, which images in sequence 312 are taken at the same time as which images in sequence 314.
• the present invention provides a system and technique for determining the correct relationship between the images 308 and 310 and more generally the correct spatial and temporal relationships between sequences 312 and 314, as shown in Fig. 3D.
• the present invention determines which image in sequence 312 corresponds in time with which image in sequence 314 and further determines the spatial relationship between the images which correspond in time.
• image A (310) in sequence 314 is found to correspond in time with image a (308) in sequence 312.
• the correct spatial relationship between images A (310) and a (308) is shown in Fig. 3D at reference numeral 350.
• image B (330) in sequence 314 is found to correspond in time with image b (322) in sequence 312
• image C (332) in sequence 314 is found to correspond in time with image c (324) in sequence 312
• image D (334) in sequence 314 is found to correspond in time with image d (326) in sequence 312.
• the correct spatial relationship between images B (330) and b (322) is shown in Fig. 3D at reference numeral 352
• the correct spatial relationship between images C (332) and c (324) is shown in Fig. 3D at reference numeral 354
• the correct spatial relationship between images D (334) and d (326) is shown in Fig. 3D at reference numeral 356.
• the present invention employs an appreciation that despite there not necessarily being any common static visual information in the two sequences, the image to image dynamics in each of the two sequences are nevertheless related both spatially and temporally.
• the present invention utilizes the relationship between the dynamics to correlate the two sequences in time and space, without requiring understanding or interpretation of the features captured by the sequences.
• the image to image dynamics preferably are expressed by spatial transformations over time between images in each sequence.
• the problem of aligning sequences of images in time and space is thus reduced to a problem of determining the relationship between sequences of such transformations in time and in space.
• FIGs. 4A - 4D illustrate resolution of spatial and temporal relationships between sequences of images of a scene taken at two different times, producing two corresponding sequences of images.
• a camera used for producing the sequences of images is fixed to a element which travels along a path that is generally identical for both sequences.
• a camera mounted onto an element traveling along a generally identical path, such as a railroad track, images a scene at two different times, ' here typically summer and winter.
• the camera used in the summer is designated by reference numeral 400
• the camera used in the winter is designated by reference numeral 402
• the same camera or different cameras may be employed.
• the cameras may have the same or different settings.
• Each camera takes a sequence of images of the scene.
• the scene being imaged at the two different times has two different appearances, as designated by reference numerals 404 and 406.
• Reference numeral 404 designates a view of a train bearing camera 400, traveling along railroad track 407 in the summer.
• Reference numeral 406 designates a view " of a train bearing camera 402 traveling along railroad track 407 in the winter.
• Fig. 4B cameras 400 and 402 image portions of the views designated by reference numerals 404 and 406 of Fig. 4A, generating respective images 408 -and 410.
• images 408 and 410 each belong to a sequence of images, each produced by one of the moving cameras 400 and 402, the respective sequences being designated by reference numerals 412 and 414.
• sequence 412 also includes images 422, 424 and 426.
• sequence 414 also includes images 430, 432 and 434. It is appreciated that the images in sequence 412 are not necessarily spaced in time (taken at the same time differences) as the images in sequence 414.
• features imaged in the summer may be different from the features imaged in the winter due, for example to snow cover.
• a house seen in sequence 412 is not visible in sequence 414, because it is covered by snow.
• snow seen in sequence 414 is not visible in sequence 412.
• a problem addressed by the present invention is that the visual information contained in individual pairs of images, one belonging to sequence 412 and the other belonging to sequence 414, (e.g. (408, 410), (422, 430) or (422, 434)), is sufficient to establish neither the spatial nor the temporal relationship between two images of a pair. More generally, the visual information contained in individual pairs of images, one of the pair belonging to sequence 412 and the other of the pair belonging to sequence 414, is sufficient to establish neither the spatial nor the temporal relationship between the two sequences 412 and 414.
• two images are the to be in the same "temporal relationship" if an element to which the camera is fixed is at the same location along the path. It is appreciated that the time differences between images 422, 424, 408 and 426 in sequence 412 may or may not be the same as the time differences between the images 430, 432, 410 and 434 in sequence 414.
• the unknown spatial relationship of images 408 and 410 is seen graphically by considering three examples of possible relative spatial relationships shown in Fig. 4B and designated by reference numerals 416, 418 and 420. In each example, the two images 408 and 410 are placed in a different spatial relationship, all of which are consistent with the visual content of the images 408 and 410.
• sequences 412 and 414 are seen graphically by considering Fig. 4C. It is appreciated that it is not apparent from a cursory examination of sequences 412 and 414, which images in sequence 412 are taken from the same position along the railroad track 407 as images in sequences 414.
• the present invention provides a system and technique for determining the correct relationship between the images 408 and 410 and more generally the correct spatial and temporal relationships between sequences 412 and 414, as shown in Fig. 4D.
• the present invention determines . which image in sequence 412 corresponds temporally with which image in sequence 414 and further determines the spatial relationship between the images which correspond temporally.
• image A (422) in sequence 412 is found to correspond temporally with image a (430). in sequence 414.
• the correct spatial relationship between images A (422) and a (430) is shown in Fig. 4D at reference numeral 450.
• image B (424) in sequence 412 is found to correspond temporally with image b (432) in sequence 414
• image C (408) in sequence 412 is found to correspond temporally with image c (410) in sequence 414
• image D (426) in sequence 412 is found to correspond temporally with image d (434) in sequence 414.
• Fig. 4D The correct spatial relationship between images B (424) and b (432) is shown in Fig. 4D at reference numeral 452, the correct spatial relationship between images C (408) and c (410) is shown in Fig. 4D at reference numeral 454 and the correct spatial relationship between images D (426) and d (434) is shown in Fig. 4D at reference numeral 456.
• the location of the house buried under the snow may be determined in sequence 414 although the house is -not visible in that sequence.
• the present invention employs an appreciation that despite there not necessarily being any common static visual information in the two sequences, the image to image dynamics in each of the two sequences are nevertheless related both spatially and temporally.
• the present invention utilizes the relationship between the dynamics to correlate the two sequences spatially and temporally, without requiring understanding or interpretation of the features captured by the sequences.
• the image to image dynamics preferably are expressed by spatial transformations over time between images in each sequence.
• the problem of aligning sequences of images spatially and temporally is thus reduced to a problem of determining the relationship between sequences of such transformations.
• aligned sequences may be realized, which sequences may then be employed to determine the location of objects visible in only some sequences across other sequences.
• the aligned sequences may also be employed to detect changes in a scene, such as man-made changes, which take place during a time interval between acquisitions of different sequences.
• the spatial relationship may also be produced by moving at least one camera at different times and/or in different locations, along generally different trajectories, wherein the at least two trajectories are correlated as if they were produced by two cameras that are mounted rigidly on the same device or rig, and moved jointly in space. It is appreciated that Figs.
• the output can be numerical, in the form of spatial and temporal transformations, or the output can be visual in the form of aligned and/or integrated ' video sequences.
• the output transformations can be provided in various possible coordinate systems and when visual outputs are provided, there are many possible visualizations.
• Figs. 5 A - 5C illustrate an important inventive principle underlying the present invention.
• a difficulty exists in establishing spatial and temporal correspondence between images that have little or no visual information in common.
• the present invention overcomes this difficulty by considering intra-sequence spatial transformations between images in each sequence and by then correlating series of such intra-sequence spatial transformations across the sequences.
• Figs. 5A - 5C and the following description present examples of these correlations which can be employed in accordance with the present invention to derive the spatial and temporal relationship, notwithstanding a lack of common visual information across the image sequences.
• Fig. 5A illustrates the relationships between image to image transformations within two sequences, induced by motion of first and second cameras 500 and 502 along an axis 504, the two cameras being arranged at 180 degrees with respect to each other along the axis.
• Camera 500 captures a sequence of images 506, while camera 502 captures a sequence of images 508.
• Sequence 506 includes inter alia images 510, 512 and 514, while sequence
• 508 includes inter alia images 520, 522 and 524. It is seen that the motion of camera 500 along axis 504 in a direction indicated by arrow 526 causes an imaged scene including a house to appear smaller in image 512 than in image 510. This relationship is represented by an intra-sequence image-to-image spatial transformation and the graphical illustration of the transformation from image 510 to image 512 is designated by reference numeral 530. The arrows shown in the transformation represent the displacement of the corresponding points of the image 512 as the image 512 is transformed to image 510. Similarly, it is seen that the motion of camera 500 along axis 504 in a direction indicated by arrow 526 causes the imaged scene including the house to appear smaller in image 514 than in image 512. This relationship is represented by an intra-sequence image-to-image spatial transformation designated by reference numeral 532.
• FIG. 5B illustrates the relationships between image to image transformations within two sequences, induced by motion of first and second cameras 550 and 552 along an axis 554, the two cameras being arranged at 90 degrees with respect to each other, one of the cameras being aligned substantially parallel to the axis of motion 554.
• Camera 550 captures a sequence of images 556, while camera 552 captures a sequence of images 558.
• Sequence 556 includes inter alia images 560, 562 and 564, while sequence 558 includes inter alia images 570, 572 and 574. It is seen that the motion of camera 550 along axis 554 in a direction indicated by arrow 576 causes an imaged scene including a house to appear shifted sideways to the left in image 562 as compared with image 560. This relationship is represented by an intra-sequence image-to-image spatial transformation designated by reference numeral 580. Similarly, it is seen that the motion of camera 550 along axis 554 in a direction indicated by arrow 576 causes the imaged scene including the house to appear shifted sideways to the left in image 564 as compared with image 562.
• This relationship is represented by an intra-sequence image-to-image spatial transformation designated by reference numeral 582. It is also seen that the motion of camera 552 along axis 554 in a direction indicated by arrow 576 causes an imaged scene including a tree to appear larger in image 572 than in image 570. This relationship is represented by an intra-sequence image-to-image spatial transformation designated by reference numeral 590. Similarly, it is seen that the motion of camera 552 along axis 554 in a direction indicated by arrow 576 causes the imaged scene including the tree to appear larger in image 574 than in image 572. This relationship is represented by an intra-sequence image-to-image spatial transformation designated by reference numeral 592.
• Fig. 5C illustrates the relationships between image to image transformations within two sequences, induced by motion of first and second cameras 600 and 602 along an axis 604, the two cameras being directed in the same direction perpendicular to the direction of motion along axis 604 but at different zooms.
• Camera 600 captures a sequence of images 606, while camera 602 captures a sequence of images 608.
• Sequence 606 includes inter alia images 610, 612 and 614, while sequence 608 includes inter alia images 620, 622 and 624. It is seen that the motion of camera 600 along axis 604 in a direction indicated by arrow 626 causes an imaged scene including a house to appear shifted sideways to the left in image 612 as compared with image 610. This relationship is represented by an intra-sequence image-to-image spatial transformation designated by reference numeral 630. Similarly, it is seen that the motion of camera 600 along axis 604 in a direction indicated by arrow 626 causes the imaged scene including the house to appear shifted sideways to the left in image 614 as compared with image 612. This relationship is represented by an intra-sequence image-to-image spatial transformation designated by reference numeral 632.
• FIG. 6A, 6B and 6C illustrate that computational functionality employing one type of transformation may not ' be sufficient for resolving ambiguities in the spatial and temporal relationships between sequences but that employing multiple different types of transformations reduces ambiguities in the spatial and temporal relationships between sequences.
• intra-sequence spatial transformations 650 and 652 representing two oppositely directed sideways movements, each belonging to a sequence taken by a different camera, can represent at least two different arrangements of the cameras.
• One such arrangement, designated by reference numeral 654 is of first and second cameras, 656 and 658 respectively, moving in a direction 660 along an axis 662 and being arranged in a mutual upside down arrangement and both directed in parallel perpendicular to axis 662, • -
• Another possible arrangement is of cameras 656 and 658 being arranged at 180 degrees with respect to each other, directed perpendicular to the axis 662 and moving along axis 662 in a direction 660 therealong.
• intra-sequence spatial transformations 670 and 672 representing forward and backward motion respective, each belonging to a sequence taken by a different camera, can represent at least two different arrangements of the cameras.
• One such arrangement, designated by reference numeral 664 " is of first and second cameras, 656 and 658 respectively, moving in a direction 680 along an axis 682 and being arranged at 180 degrees with respect to each other along axis 682.
• Another possible arrangement is of cameras 656 and 658 being arranged at 180 degrees with respect to each other along axis 682 for motion in a direction 680 therealong, wherein one of the cameras 656 is rotated about axis 682 with respect to the other one of the cameras 658.
• Fig. 6C it is seen that two series of intra-sequence spatial transformations 690 and 692, each taken by a different moving camera, and containing transformations of different types, are employed by computational functionality to resolve ambiguities of the type described hereinabove with reference to Figs. 6A and 6B.
• Series 690 comprises transformation 650 (Fig. 6 A) followed by transformation 670 (Fig. 6B), while ' series 692 comprises transformation 652 (Fig. 6A) followed by transformation 672 (Fig. 6B).
• Figs. I to 4D described hereinabove present four typical applications of the present invention.
• the present invention provides alignment between images of two sequences of images notwithstanding that there is insufficient similarity between any image of one sequence to any image of the other sequence to enable the prior art to provide image to image alignment.
• "Similarity" of images is used here in a broad sense and includes, inter alia, gray-level similarity, feature similarity, similarity of frequencies and statistical similarity, such as mutual information. Consequently, in the prior art, at least partial overlap between the images is required to obtain similarity.
• the present invention employs correlated temporal behavior between the two sequences, as shown and described with reference to Figs. 5A to 6C, to resolve both the spatial and the temporal transformations between images of the two sequences.
• a features of the present invention is the ability to replace the requirement for "coherent appearance", which is a fundamental assumption in the prior art, with the requirement for "consistent temporal behavior" which is often easier to satisfy, for example by moving the two cameras jointly in space.
• the two sequences are produced concurrently by two cameras moving jointly in space.
• the term 'concurrently' here designates that the two sequences are produced at least partially during the same period of time.
• the concurrency does not require that images, or frames, of the two sequences be produced simultaneously or even at the same rate.
• 'Moving jointly' here means that the two cameras have, for practical purposes, the same center of projection throughout the sequences. This may be realized, for example by fixedly attaching both cameras to the same moving rig. In reality, the centers of projection of the two cameras are not at precisely the same location.
• the centers of projection of both cameras are considered as if they were at the same location. It is further appreciated that the mathematical principles employed in the mathematical treatment are also applicable to situations where the centers of projection of the cameras are clearly not at the same location but are at locations whose spatial relationship is fixed. It is additionally appreciated that the mathematical principles employed in the mathematical treatment are also applicable to situations where the centers of projection of the cameras are clearly not at the same location but are at locations having an at least partially known spatial/temporal relationship.
• the two sequences contain images that are partially overlapping.
• the similarity between the images is very low or even nonexistent, namely, no detail of any image, of one sequence can be identified in any image of the second sequence.
• the two sequences are produced by cameras employing different sensing modalities.
• Typical sensing modalities include: using a photographic camera, such as a still film camera, a digital camera, a video camera or a motion film camera; employing infra-red imaging equipment, radar, x-ray, CAT-scan, MRI or other electromagnetic imaging equipment, acoustic imaging equipment, such as ultrasound, sonar and sub-acoustic geophysical surveillance equipment, satellite remote sensing equipment.
• a photographic camera such as a still film camera, a digital camera, a video camera or a motion film camera
• infra-red imaging equipment radar, x-ray, CAT-scan, MRI or other electromagnetic imaging equipment
• acoustic imaging equipment such as ultrasound, sonar and sub-acoustic geophysical surveillance equipment, satellite remote sensing equipment.
• the two sequences are produced at different times but the two cameras that produce the respective two sequences are moving generally along the same path.
• the sequences may not have any spatial
• Such lack of similarity may result from differences in internal camera calibration, such as magnification, zoom and resolution, or from employing different sensing modalities. This lack of similarity may also result from changes in the scene, such as changes caused by the seasons, as illustrated in Figs. 4A to 4D, or due to differences in visibility between the two sequences.
• the image to image dynamics preferably are expressed by spatial transformations over time between images in each sequence.
• the problem of aligning sequences of images in time and space is thus reduced to a problem of determining the relationship between sequences of such transformations in time and in space.
• the motions of the two cameras induce motions to the. two sequences.
• the induced motions may be different for the two sequences but are correlated.
• the correlation between the induced motions serves to correlate between sequences both in time and in space.
• Each of the images must be a part of a temporal sequence of images; (2) The two sequences must be produced during relative motion between cameras and a scene; (3) The optical parameters of the two cameras must not change other than in a known manner throughout the production of the two sequences.
• the present invention does not require that any single image in one sequence of images have any spatial or visual similarity with any single image in the other sequence. Consequently and in addition, the present invention provides alignment between two images belonging to two sequences of images when the level of similarity between these images is insufficient to provide alignment by employing the prior art.
• the present invention normally does not require any prior internal or external camera calibration at any time.
• Figs. 6A to 6C simple camera motion may not be sufficient to resolve the spatial and the temporal transformations between images of the two sequences. Therefore several different types of camera motions may be required. Three examples of such camera motion are shown in Figs. 5 A to 5C, it being understood that other, perhaps better, examples of camera motion may exist.
• An embodiment of the present invention described hereinbelow provides assessment of the required complexity of the motion.
• FIG. 7 is a simplified illustration of a complex motion 700 of two cameras 702 and 704, fixed to each other, each camera producing a sequence of images of a portion of a scene 703, the sequences being 706,
• Fig. 8 is a simplified illustration portions of two sequences 720 and 722 of images taken as shown in Fig. 7, wherein the two sequences are spatially related by a fixed and unknown inter-camera homography and temporally related by a fixed and unknown time shift ⁇ t.
• the symbols Ij, Ti, T'j and H are defined hereinbelow.
• Fig. 9 is a functional block diagram of a preferred sequence of steps of implementation of a preferred embodiment of the present invention.
• step 900 two cameras 910 and 920 are employed to produce two sequences of images 930 and 940.
• S be a first sequence of images I t produced by a first camera and let S' be a second sequence of images /' / produced by a second camera, for example as shown and described in accordance with any of the Figs. IA to 8 hereinabove, and wherein S The two input sequences have an unknown temporal relation and an unknown spatial relation between them.
• the spatial relation may- represent the fact that the two cameras are firmly attached to each other.
• the two cameras are mounted rigidly on the same device or rig, and moved jointly in space.
• the spatial relation may represent the fact that the cameras are moving in a generally identical trajectory but at different times.
• the spatial relation is produced by moving at least one cameras in a generally identical trajectory but in at least one of a different time and a different location, wherein there is a fixed relative position and a fixed relative orientation between the two cameras and wherein the optical . properties of each of the two cameras do not change between the images /, ⁇ and / + ⁇ and between /',- and /', ⁇ +!
• the spatial relationship is produced by moving at least one cameras in a generally identical trajectory but in at least one of a different time and a different location, wherein the relative position and a relative orientation between the two cameras may change in a known way and wherein the optical properties of each of the two cameras may change between the images I, and 7,-+ ⁇ and between /' and /', • +! m & known way.
• the spatial relationship is produced by moving at least one camera at different times and/or in different locations, along generally different trajectories, wherein the at least two trajectories are correlated as if they were produced by two cameras that are mounted rigidly on the same device or rig, and moved jointly in space.
• the temporal relation between the two sequences represents the time when the two cameras were in the spatial relation as described hereinabove.
• the two sequences are produced by . employing synchronization equipment to synchronize between the two cameras.
• the temporal relation between the two sequences represents the fact that the image /,- and the corresponding image /',- are produced together.
• the temporal relation between the two sequences represents the fact the image /, ⁇ and the corresponding image /',- are produced with a fixed time delay between them.
• the temporal relation applies to image /,• and a point in sequence S' that is between a corresponding /',- and image /'/+ ⁇ -
• the temporal relation applies to two sequences produced employing two cameras employing two different frame rates.
• the temporal relation between the two sequences may represent the fact that the two images are the to be in the same "temporal relationship" if the cameras producing the sequences were in the same location In the same time.
• temporal relation between sequences may exist even if the temporal relation between images may hold only for virtual images, that is, for images that have not been produced but could have been produced if the camera was producing an image at that time.
• S and S' are not necessarily the original sequences, but may by subsequences of the original sequences, where subsequence may represent, a portion of the field-of-view, temporal portion of the sequence, sub-sampling of pixels, and sub-sampling of frames.
• sequences S and S' may be produced in several methods as described hereinbelow:
• the first and the second camera may have a different 3D Orientation and move jointly in space, such as the camera pairs shown in Fig 7.
• none of the images / contains any feature contained in any of the images /'hom however, there is at least a partial temporal overlap between S and S '.
• the first camera and the second camera have at least partially overlapping fields of view, wherein images
• / comprises details that are not comprised in the overlapping parts of the corresponding images V ⁇ and wherein the two cameras move jointly in space, such as the camera pair shown and described in accordance with Figs. 3A to 3D hereinabove.
• sequence S and the sequence S ' may be produced by the same camera moving in at least approximately the same pattern in space and at different times.
• sequence S and the sequence S' may be produced by two cameras moving in at least approximately the same pattern in space and at different times.
• sequences S and S' may be produced by any image producing instrument that produces at least two dimensional images of space, such as photographic camera, such as still film camera, digital camera, video camera or motion film camera, infra-red imaging equipment, radar, x-ray, CAT-scan, MRI and other electromagnetic imaging equipment, acoustic imaging equipment, such as ultrasound, sonar and sub-acoustic geophysical surveillance equipment and satellite remote sensing equipment. It is also appreciated that the two sequences S and S' may be produced by image producing instruments of the same type or of different types. It is appreciated that neither of the two sensors has to be calibrated.
• the temporal correspondence can be achieved electronically, for example by employing Genlock, or recovered using the method describe hereinbelow with reference to element 960 of Fig. 9.
• a method for computing temporal relation described hereinbelow can compute temporal relation at sub-frame time units.
• a temporal relation is computed between a transformation T,- and a transformation Tj' +s wherein T / ⁇ s is interpolated from 7 ⁇ - and 7 ⁇ + ⁇ .
• the sequence of transformations 7 ⁇ +( y applies to virtual images I /+ s, wherein in the virtual images were not actually produced by the camera producing image sequence S', but could have been produced by the camera if the images produced at the appropriate time points and are therefore interpolated representations of the scene.
• step 990 of Fig. 9. It can now be assumed that the transformations and images are temporally synchronized.
• H be a spatial transformation function that transforms between a coordinate system of sequence S and a coordinate system of sequence S'. Applying the transformation H on the sequence S' results in the content of images /',• in the coordinates of the sequence S, as if the transformed sequence S' has been produced by the camera that has produced the sequence S. Therefore:
• p ' H(p) where p ' is the feature p of /,- in the coordinates of /',-.
• H is a 3 x 3 inevitable matrix.
• H is a 4 x 4 inevitable Euclidean or inevitable projective matrix.
• the two sequences S and S do not share sufficient common feature or any other properties of the images. Therefore we cannot compute H from common properties of the images / / . Instead H is computed by correlating between properties of the temporal progressions within each sequence
• H is computed from the induced frame to frame transformation within each of the sequences S and S'.
• a large portion of the scene that is photographed by the two cameras, producing the sequences S and S' is planar or distant from the cameras.
• J; and T ave 3x3 non-singular matrices.
• P be a three dimensional point in the planar or distant scene and let p,- an ⁇ p ', be the homogeneous coordinates of the projection of the point P in the images /; and I) respectively. It is noted that there is no need for P to appear in both or in any of the images /,- and I), that is, within the field of view of the cameras when images /,- or-E; are produced.
• Let ?,- +! and /? ', • +! be the coordinates of the projection of the point P in images I/+ ⁇ and /'/+] respectively.
• Equation (4) implies that eig(T ⁇ and eig(T',) axe "parallel". This gives rise to a measure of similarity between two matrices T, and T' constitute denoted by sim(T réelle T',) and presented in Eq. (5) hereinbelow:
• sim(T,, T ',) vector norm .
• Eq. (4) is a set of three equations with one unknown and can be solved using least squares minimization.
• the input homographies 77 and T can be normalized to have determinant equal 1 and to avoid the need to compute s,.
• Equation (6) is linear in the unknown components of H. Rearranging the components of H in a 9 x 1 column vector h - [H ⁇ H ⁇ 2 Hi3H2iH22H23H3iH32H33] t 5 Eq. (6) can be rewritten as a set of linear equations in h in the form of Eq. (7) hereinbelow:
• ,- is a 9 x 9 matrix defined by Tj, T and s.
• Equation (8) ' is a homogeneous set of line equations in h that can be solved in a variety of ways.
• h is solved by computing the eigenvector which corresponds to the smallest eigenvalue of the matrix A' A.
• the scene for which the sequences S and S' have been produced is neither planar nor distant.
• the temporal progression between any two consecutive images /,- and I i+ ⁇ is described by a fundamental matrix E,-.
• the fundamental matrix F defines the relation between corresponding image points ', so that:
• F F ⁇ ,..,F n
• F F ⁇ ,..,F n
• H is computed from properties of the temporal progression output as expressed by the sequences F and F' of the matrices E,- andF .
• Each fundamental matrix E,- can be decomposed into a homography + epipole as described by Eq. (9) hereinbelow::
• the homographies, T ⁇ ,...,T n and T ,...,T' n , and the epipoles e ⁇ ,...,e n and e impose separate constraints on the inter-camera homography H. These constraints can be used separately or jointly to recover H as described hereinbelow.
• the homographies T ⁇ ,...,Tmony and T ,...,T' n extracted from the fundamental matrices F ⁇ ,...,Fspawn and F' ⁇ ,...,F' n , respectively may correspond to different three dimensional planes.
• the "Plane+Parallax” method which is well known in the prior art, is used to impose plane-consistency across and within the two sequences T ⁇ ,...,T n and T' ⁇ ,...,T' scenery.
• the Plane+Parallax method describes the image-to-image transformation using two components: homography and residual parallax.
• the homography is associated with a selected plane and relates points on this plane between the two images.
• the parallax component describes the residual displacement of off-plane points.
• the "Plane+Parallax” method requires that a real physical planar surface be visible in all images 7,- and .
• One method for computing a Plane+Parallax transformation is taught by M.
• the homography part of the Plane+Parallax transformation is then used in the same manner as in the planar or distant case described hereinabove.
• the "threading" method as described for example in S. Avidan and A. Shashua. Threading fundamental matrices. In European Conference on Computer Vision, 1998, which is also well known in the prior- art, is used to impose plane-consistency across and within the two sequences T],...,T n and T' ⁇ ,...,T'êt . .
• the "threading" method can impose plane-consistency within each sequence, even if no real physical plane is visible in any of the frames.
• camera matrix refers to a 4 x 3 matrix, which is comprised of a 3 x 3 left minor matrix and a 1 x 3. vector, wherein the 3 x 3 left minor matrix represents a 3 x 3 homography.
• the consistency of the camera matrices imposes plane-consistency on the 3 x 3 homographies.
• plane consistency between the two sequences is provided by applying the threading method on each of the two sequences S and S', by initiating the threading method at frames which are known to simultaneously view the same real plane in both sequences. It is appreciated that the two cameras can see different portions of the plane, allowing for non-overlapping fields of view, and need not see the plane at any of the other images.
• the homography H is computed from epipoles e, and e 7, which are computed from the fundamental matrices ⁇ F; and E7 Given two images / ,- and I i+ ⁇ , an epipole denotes the spatial location of the center of projection of the camera for image 7, + ⁇ in the coordinates system of image I.
• the fundamental matrices provide a list of epipoles e ⁇ ,...,e n and e ,...,e ' note, wherein an epipole ⁇ z,- is the null ' space of a fundamental matrice F,- and an epipole e is the null space of a fundamental matrice F' ⁇ .
• the epipoles e ,- and e- can be computed by other methods than from the null space of the fundamental matrices F,-. and F'u respectively. There are many other methods known in the prior art to find epipoles between two images.
• h is solved by computing the eigenvector which corresponds to the smallest eigenvalue. Theoretically, four pairs of corresponding epipoles e ⁇ - and e 7 are sufficient, provided that no three epipoles are on the same line.
• H can be solved by adding the system of 2n equations to the set of linear equations in Eq. (8) which are imposed by the homographies.
• the image 7 in sequence S corresponds to image 77+ ⁇ t m sequence S' and ⁇ t is unknown. Therefore the transformation T, corresponds to transformation T and not to T . If time stamping, or similar additional information, is available for each image 7,- and 77 then synchronization can be recovered. When there is no additional information to recover the synchronization, the synchronization can be recovered in using the method hereinbelow.
• T ,...,T' m wherein T, and T i+ ⁇ t are temporally corresponding transformations.
• the maximization is performed by an exhaustive search over a finite range of valid time shifts ⁇ t.
• Coarser temporal levels are constructed by composing transformations to obtain fewer transformation between more distant frames.
• the images in sequence S are produced at a different rate than the images of sequence S '.
• sequence S is produced by an NTSC camera at 30 frame per second and sequence S' is produced by a PAL camera at 25 frames per second.
• T. J is the transformation from frame I, to frame I j .
• PAL and NTSC video cameras produce each image as pair of fields, wherein the first field comprises the odd- lines of the image and the second field comprises the even lines of the picture.
• the fields are produced in a rate that is double the rate of the images, or frames, that is at 60 and 50 fields per second respectively.
• the method described above with respect to Eq. (12. a) can recover synchronization to the field level.
• Sub-field accuracy can be recovered by interpolating the values of SIM( At) obtained at discrete time shifts.
• the transformations are interpolated according to the computed temporal correspondence and SIM( ⁇ t) is re-estimated, in an iterative manner, to increase the temporal accuracy.
• the number N m ⁇ - n is assessed by examining via the similarity equation Eq. (3) the number of constraints imposed on Hby a single pair of transformations 7 and T .
• Similar matrices also known as conjugate matrices, such as B and G, have the same eigenvalues but different eigenvectors.
• eigen subspaces wherein if V is an eigen- subspace of B corresponding to an eigenvalue ⁇ , then H(V) is an eigen subspace G with the same eigenvalue ⁇ .
• This case typically occurs in one of the following two cases: (a) when there exists an eigenvalue with algebraic multiplicity two; or (b) when there is only one eigenvalue with algebraic multiplicity three,
• the corresponding eigen subspaces has dimensionality 2. It is spanned by two linearly independent eigenvetors [1,0,0]' and [0,1,0]'.
• each of the two eigenvectors u b ⁇ and u b2 provides 3 linear equations and 2 new unknowns. Therefore, together the two eigenvectors U b i and u b2 provide 2 new linear constraints on H.
• step 950 a frame-to-frame input transformations, T ⁇ ,...,T n and
• Step 970 comprises methods of robust statistics to enhance the accuracy of the transformations 7 and 777-.
• step 970 comprises two optional methods: (a) Outlier rejection, designated by reference numeral 972; and
• step 972 inaccurate frame-to-frame transformations 77,- are preferably pruned out by employing any of the two outlier detection mechanisms known in the prior art as described hereinbelow:
• the transformation between successive frames within each sequence are computed in both directions.
• the distance of the composed matrix f r ⁇ m the identity matrix in the image space is computed, in terms of the maximal residual misalignment of pixels.
• step 974 optionally, any one of two methods of coordinate renormalization known in the prior art is used as described hereinbelow:
• Input matrices are normalized so that the rows of M has approximately the same norm, using the heuristic provided in for example, Gene Golub and Charles Nan Loan. Matrix Computations. The Johns Hopkins University Press, Baltimore and London, pp. 123-127, 1989.
• temporal sub-sampling of the sequences can provide more significant transformations between successive frames.
• temporal sub-sampling is done after recovering the temporal synchronization, in order to guaranty that temporally corresponding frames are sampled from the two video sequences.
• the three dimensional joint motions of the two cameras are planned in advance to increase the accuracy of the computed result.
• input data are sequences of transformation and T'i.
• the transformation 77 and 77- are extracted from two video sequences using one of a variety of algorithms provided in the prior art.
• the reliability of the input data also depends on the specific algorithm used to extract the input transformations.
• Employing a rough approximation to the accuracy of the different ' transformations Tj and T it is possible to analyze the combinations of the transformations 77- and T that participate in the matrix M (EqJ).
• EqJ matrix M

Landscapes

• Engineering & Computer Science (AREA)
• Computer Vision & Pattern Recognition (AREA)
• Physics & Mathematics (AREA)
• General Physics & Mathematics (AREA)
• Theoretical Computer Science (AREA)
• Multimedia (AREA)
• Image Analysis (AREA)
• Image Processing (AREA)
• Measuring And Recording Apparatus For Diagnosis (AREA)
• Apparatus For Radiation Diagnosis (AREA)
• Studio Circuits (AREA)

Applications Claiming Priority (3)

Publications (1)

Family

ID=22951289

Family Applications (1)

Country Status (7)

Families Citing this family (123)

Family Cites Families (27)

• 2001
  • 2001-12-04 WO PCT/IL2001/001120 patent/WO2002047031A2/en active Application Filing
  • 2001-12-04 AU AU2002221023A patent/AU2002221023A1/en not_active Abandoned
  • 2001-12-04 US US10/433,526 patent/US7596281B2/en not_active Expired - Fee Related
  • 2001-12-04 CA CA002436607A patent/CA2436607A1/en not_active Abandoned
  • 2001-12-04 EP EP01999922A patent/EP1356431A2/de not_active Withdrawn
  • 2001-12-04 IL IL15625001A patent/IL156250A0/xx not_active IP Right Cessation
  • 2001-12-04 JP JP2002548680A patent/JP2004523027A/ja active Pending

Non-Patent Citations (1)

Also Published As

Similar Documents

Legal Events